search

alan-turing-institute / grace

Graph Representation Analysis for Connected Embeddings
Other
34 stars 1 forks source link

What's the reasoning behind the dropout positioning? #181

Open KristinaUlicna opened 1 year ago

KristinaUlicna commented 1 year ago

I'm struggling to understand why we're putting the dropout here & only here at train time:

https://github.com/alan-turing-institute/grace/blob/e634d3621747d975f974aade96141bb9174d5594/grace/models/classifier.py#L71-L73

Any suggestions how this could be experimented with?

chris-soelistyo commented 1 year ago

As far as I can tell, no ironclad reason for this; as an experiment could also try placing it after each graph convolutional layer.